Cleanup NIXL file mounts

src/cloudai/_core/command_gen_strategy.py

@@ -1,5 +1,5 @@
 # SPDX-FileCopyrightText: NVIDIA CORPORATION & AFFILIATES
-# Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -49,6 +49,10 @@ def store_test_run(self) -> None:
         """
         pass
 
+    def cleanup_job_artifacts(self) -> None:
+        """Best-effort cleanup hook run after the job has fully completed."""
+        return
+
     @property
     def final_env_vars(self) -> dict[str, str | list[str]]:
         if not self._final_env_vars:

src/cloudai/systems/slurm/single_sbatch_runner.py

@@ -22,7 +22,7 @@
 from typing import Generator, Optional, cast
 
 from cloudai.configurator import CloudAIGymEnv, TrajectoryEntry
-from cloudai.core import JobIdRetrievalError, System, TestRun, TestScenario
+from cloudai.core import BaseJob, JobIdRetrievalError, System, TestRun, TestScenario
 from cloudai.util import CommandShell, format_time_limit, parse_time_limit
 
 from .slurm_command_gen_strategy import SlurmCommandGenStrategy
@@ -221,6 +221,9 @@ def handle_dse(self):
                     )
                 )
 
+    def completed_test_runs(self, job: BaseJob) -> list[TestRun]:
+        return list(self.all_trs)
+
     def _submit_test(self, tr: TestRun) -> SlurmJob:
         with open(self.scenario_root / "cloudai_sbatch_script.sh", "w") as f:
             f.write(self.gen_sbatch_content())

src/cloudai/systems/slurm/slurm_runner.py

@@ -1,5 +1,5 @@
 # SPDX-FileCopyrightText: NVIDIA CORPORATION & AFFILIATES
-# Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -77,10 +77,18 @@ def on_job_submit(self, tr: TestRun) -> None:
         cmd_gen = self.get_cmd_gen_strategy(self.system, tr)
         cmd_gen.store_test_run()
 
+    def completed_test_runs(self, job: BaseJob) -> list[TestRun]:
+        return [cast(SlurmJob, job).test_run]
+
     def on_job_completion(self, job: BaseJob) -> None:
         logging.debug(f"Job completion callback for job {job.id}")
         self.system.complete_job(cast(SlurmJob, job))
         self.store_job_metadata(cast(SlurmJob, job))
+        for tr in self.completed_test_runs(job):
+            try:
+                self.get_cmd_gen_strategy(self.system, tr).cleanup_job_artifacts()
+            except Exception:
+                logging.warning(f"Cleanup failed for test run at {tr.output_path}", exc_info=True)
 
     def _mock_job_metadata(self) -> SlurmStepMetadata:
         return SlurmStepMetadata(

src/cloudai/workloads/common/nixl.py

@@ -17,6 +17,7 @@
 
 import logging
 import re
+import shutil
 from functools import cache
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Final, Generic, TypeVar, cast
@@ -231,6 +232,25 @@ def _unique_file_name(self, file_name: str, used_filenames: set[str]) -> str:
         used_filenames.add(candidate)
         return candidate
 
+    def cleanup_job_artifacts(self) -> None:
+        for cleanup_target in self._cleanup_targets():
+            if cleanup_target.is_dir():
+                shutil.rmtree(cleanup_target)
+                logging.debug(f"Cleaned up job artifact: {cleanup_target}")
+
+    def _cleanup_targets(self) -> list[Path]:
+        cleanup_targets: list[Path] = []
+
+        filepath_raw: str | None = cast(str | None, self.test_run.test.cmd_args_dict.get("filepath"))
+        if filepath_raw:
+            cleanup_targets.append(self.test_run.output_path / "filepath_mount")
+
+        device_list_raw: str | None = cast(str | None, self.test_run.test.cmd_args_dict.get("device_list"))
+        if device_list_raw and get_files_from_device_list(device_list_raw):
+            cleanup_targets.append(self.test_run.output_path / "device_list_mounts")
+
+        return cleanup_targets
+
     @property
     def final_env_vars(self) -> dict[str, str | list[str]]:
         env_vars = super().final_env_vars

tests/test_get_job_id.py

@@ -16,14 +16,14 @@
 
 import subprocess
 from pathlib import Path
-from unittest.mock import Mock
+from unittest.mock import Mock, patch
 
 import pytest
 
 from cloudai.core import JobIdRetrievalError, TestRun, TestScenario
 from cloudai.systems.lsf.lsf_runner import LSFRunner
 from cloudai.systems.lsf.lsf_system import LSFSystem
-from cloudai.systems.slurm import SlurmRunner, SlurmSystem
+from cloudai.systems.slurm import SlurmJob, SlurmRunner, SlurmSystem
 from cloudai.util import CommandShell
 from cloudai.workloads.sleep.sleep import SleepCmdArgs, SleepTestDefinition
 
@@ -88,6 +88,21 @@ def test_slurm_get_job_id(slurm_runner: SlurmRunner, stdout: str, stderr: str, e
     assert res == expected_job_id
 
 
+def test_slurm_runner_on_job_completion_calls_cleanup(slurm_runner: SlurmRunner):
+    tr = slurm_runner.test_scenario.test_runs[0]
+    job = SlurmJob(tr, id=1)
+    slurm_runner.store_job_metadata = Mock()
+    cleanup = Mock()
+    slurm_runner.get_cmd_gen_strategy = Mock(return_value=Mock(cleanup_job_artifacts=cleanup))
+
+    with patch.object(SlurmSystem, "complete_job") as complete_job:
+        slurm_runner.on_job_completion(job)
+
+    complete_job.assert_called_once_with(job)
+    slurm_runner.store_job_metadata.assert_called_once_with(job)
+    cleanup.assert_called_once()
+
+
 @pytest.mark.parametrize(
     "stdout, stderr, expected_job_id",
     [

tests/test_single_sbatch_runner.py

@@ -16,8 +16,9 @@
 
 import copy
 import re
+from pathlib import Path
 from typing import Generator, Optional, cast
-from unittest.mock import Mock
+from unittest.mock import Mock, patch
 
 import pandas as pd
 import pytest
@@ -506,6 +507,30 @@ def test_store_job_metadata(nccl_tr: TestRun, slurm_system: SlurmSystem) -> None
     assert sjm == SlurmJobMetadata.model_validate(toml.loads(toml.dumps(sjm.model_dump())))
 
 
+def test_on_job_completion_cleans_all_effective_test_runs(
+    dse_tr: TestRun, nccl_tr: TestRun, slurm_system: SlurmSystem
+) -> None:
+    tc = TestScenario(name="tc", test_runs=[dse_tr, nccl_tr])
+    runner = SingleSbatchRunner(mode="run", system=slurm_system, test_scenario=tc, output_path=slurm_system.output_path)
+    runner.mode = "dry-run"
+    runner.store_job_metadata = Mock()
+
+    cleanup_calls: list[Path] = []
+
+    def _cmd_gen(_, tr: TestRun):
+        return Mock(cleanup_job_artifacts=Mock(side_effect=lambda: cleanup_calls.append(tr.output_path)))
+
+    runner.get_cmd_gen_strategy = Mock(side_effect=_cmd_gen)
+
+    expected_paths = [tr.output_path for tr in runner.all_trs]
+    job = SlurmJob(nccl_tr, id=1)
+
+    with patch.object(SlurmSystem, "complete_job"):
+        runner.on_job_completion(job)
+
+    assert cleanup_calls == expected_paths
+
+
 def test_pre_test(nccl_tr: TestRun, sleep_tr: TestRun, slurm_system: SlurmSystem) -> None:
     nccl_tr.pre_test = TestScenario(name="pre_test", test_runs=[sleep_tr])
     tc = TestScenario(name="tc", test_runs=[nccl_tr])

tests/workloads/nixl_bench/test_command_gen_strategy_slurm.py

@@ -105,6 +105,32 @@ def test_container_mounts(self, nixl_bench_tr: TestRun, slurm_system: SlurmSyste
             assert (nixl_bench_tr.output_path / "device_list_mounts" / local_device_filename).is_file()
             assert (nixl_bench_tr.output_path / "device_list_mounts" / local_device_filename).stat().st_size == 1024
 
+    def test_cleanup_job_artifacts(self, nixl_bench_tr: TestRun, slurm_system: SlurmSystem):
+        nixl_bench_tr.test.cmd_args = NIXLBenchCmdArgs.model_validate(
+            {
+                "docker_image_url": "docker.io/library/ubuntu:22.04",
+                "path_to_benchmark": "/nixlbench",
+                "backend": "GUSLI",
+                "device_list": "11:K:/dev/nvme0n1,12:F:/p1/store0.bin,13:F:/p2/store0.bin",
+                "filepath": "/data",
+            }
+        )
+        strategy = NIXLBenchSlurmCommandGenStrategy(slurm_system, nixl_bench_tr)
+        filepath_dir = nixl_bench_tr.output_path / "filepath_mount"
+        device_list_dir = nixl_bench_tr.output_path / "device_list_mounts"
+        other_file = nixl_bench_tr.output_path / "keep.txt"
+        filepath_dir.mkdir(parents=True, exist_ok=True)
+        device_list_dir.mkdir(parents=True, exist_ok=True)
+        (filepath_dir / "a.txt").write_text("x")
+        (device_list_dir / "b.txt").write_text("x")
+        other_file.write_text("keep")
+
+        strategy.cleanup_job_artifacts()
+
+        assert not filepath_dir.exists()
+        assert not device_list_dir.exists()
+        assert other_file.exists()
+
     @pytest.mark.parametrize(
         ("override", "expected_error_match", "expected_total_buffer_size"),
         (

tests/workloads/nixl_kvbench/test_command_gen_slurm.py

@@ -38,8 +38,10 @@ def kvbench() -> NIXLKVBenchTestDefinition:
 
 
 @pytest.fixture
-def kvbench_tr(kvbench: NIXLKVBenchTestDefinition) -> TestRun:
-    return TestRun(name="nixl-bench", num_nodes=2, nodes=[], test=kvbench)
+def kvbench_tr(kvbench: NIXLKVBenchTestDefinition, tmp_path) -> TestRun:
+    output_path = tmp_path / "nixl-kvbench"
+    output_path.mkdir(parents=True, exist_ok=True)
+    return TestRun(name="nixl-bench", num_nodes=2, nodes=[], test=kvbench, output_path=output_path)
 
 
 def test_gen_kvbench_ucx(kvbench_tr: TestRun, slurm_system: SlurmSystem):
@@ -124,3 +126,29 @@ def test_get_etcd_srun_command_with_etcd_image(kvbench_tr: TestRun, slurm_system
     cmd = " ".join(strategy.gen_etcd_srun_command(tdef.cmd_args.etcd_path))
     assert tdef.etcd_image is not None
     assert f"--container-image={tdef.etcd_image.installed_path}" in cmd
+
+
+def test_kvbench_cleanup_job_artifacts(kvbench_tr: TestRun, slurm_system: SlurmSystem):
+    kvbench_tr.test.cmd_args = NIXLKVBenchCmdArgs.model_validate(
+        {
+            "docker_image_url": "docker://image/url",
+            "backend": "GUSLI",
+            "filepath": "/data",
+            "device_list": "11:F:/store0.bin",
+        }
+    )
+    strategy = NIXLKVBenchSlurmCommandGenStrategy(slurm_system, kvbench_tr)
+    filepath_dir = kvbench_tr.output_path / "filepath_mount"
+    device_list_dir = kvbench_tr.output_path / "device_list_mounts"
+    other_file = kvbench_tr.output_path / "keep.txt"
+    filepath_dir.mkdir(parents=True, exist_ok=True)
+    device_list_dir.mkdir(parents=True, exist_ok=True)
+    (filepath_dir / "a.txt").write_text("x")
+    (device_list_dir / "b.txt").write_text("x")
+    other_file.write_text("keep")
+
+    strategy.cleanup_job_artifacts()
+
+    assert not filepath_dir.exists()
+    assert not device_list_dir.exists()
+    assert other_file.exists()